The Quadrilateral Loss: Additivity as a Measurable Behavior of Dense Neural Networks
A new penalty measures when dense neural networks are behaving additively, instead of forcing additivity into the architecture.
Antonio Di Cecco introduces “quadrilateral loss,” a differentiable test for feature interaction based on mixed differences between paired training points. The paper argues that many learned interactions can be removed with little cost, and that moderate regularization can improve both accuracy and additivity on small datasets. It also says pre-regularization interaction size is a weak guide to what a regularized model will keep, challenging post-hoc interaction rankings. ArXiv · AI/CL/LG's note
Antonio Di Cecco introduces “quadrilateral loss,” a differentiable test for feature interaction based on mixed differences between paired training points. The paper argues that many learned interactions can be removed with little cost, and that moderate regularization can improve both accuracy and additivity on small datasets. It also says pre-regularization interaction size is a weak guide to what a regularized model will keep, challenging post-hoc interaction rankings. ArXiv · AI/CL/LG's note
score 4